DS-YOLOv8 for Detecting Small Objects in Remote Sensing Imagery

Due to the intricate backgrounds and the multi-scale nature of the objects, detecting small objects in remote sensing images remains challenging. In order to enhance the accuracy of detection, a model called dynamically select convolution kernel, spatial boundary attention YOLO (DS-YOLOv8) is proposed. This model substitutes the original C2f module by the C2f_DynamicConv module and incorporates the SBA module into the neck. The C2f_DynamicConv module can dynamically select convolution kernel to extract richer feature information. And the SBA module can also enhance the interactive capabilities between deep features and shallow features to improve the effect of multi-scale fusion. Through experiments conducted on the NWPU_VHR_10 dataset with data augmentation, DS-YOLOv8m model improves accuracy by 3.7%, regression by 2.4%, mAP50 by 2.6%, and mAP50:95 by 1.2% compared with YOLOv8m. This indicates a noticeable improvement in the accuracy of this model.